Identification of Colon Cancer Using Multi-Scale Feature Fusion Convolutional Neural Network Based on Shearlet Transform

Meiyan Liang,Bo Li,Jiamiao Yang,Wenxiang Feng,Zhuyun Ren

doi:10.1109/access.2020.3038764

Abstract

Colon cancer identification is of great significance in medical diagnosis. Real-time, objective and accurate inspection results will facilitate medical professionals to explore symptomatic treatment promptly. However, the existing methods depend on hand-crafted features which require extensive professional expertise and long inspection period. Therefore, we propose a multi-scale feature fusion convolutional neural network (MFF-CNN) based on shearlet transform to identify histopathological image of colon cancer. The characteristic of the framework is the shearlet coefficients of histopathological image in multiple decomposition scales were extracted as supplementary feature which were also fed to the network with the original pathological image. After feature learning and feature fusion, the MFF-CNN based on shearlet transform can achieve the identification accuracy of 96% and average F-1 score of 0.9594 for colorectal adenocarcinoma epithelium (TUM) and normal colon mucosa (NORM). The false negative rate and false positive rate can be reduced to 5.5% and 2.5%, respectively. The superior performance of the network opens a new perspectives for real-time, objective and accurate diagnosis of cancer.

Highlights

Colon cancer is one of the highest mortality diseases which can lead to human death
Cruz-Roa et al.[15] use three-layer convolutional neural network (CNN) to diagnose invasive ductal carcinoma of breast cancer; the result shows that the accuracy can be improved 6% comparing with hand-crafted feature extraction
The multi-scale feature fusion convolutional neural network (MFF-CNN) based on shearlet transform can realizes automatic feature extraction, feature fusion and classification for colon cancer, which have an accuracy of 96%

Summary

INTRODUCTION

Colon cancer is one of the highest mortality diseases which can lead to human death. According to WHO global survey in 2018, there are about 1.8 million people have been diagnosed with colon cancer and its mortality rate is over 40%[1, 2]. Cruz-Roa et al.[15] use three-layer convolutional neural network (CNN) to diagnose invasive ductal carcinoma of breast cancer; the result shows that the accuracy can be improved 6% comparing with hand-crafted feature extraction. Hamad etal.[18] present a two-stage deep learning pipeline which combines fully convolutional regression network and CNN together to detect and classify the cell nuclei in colon pathology images, respectively. We propose a multi-scale feature fusion convolutional neural network (MFF-CNN) based on shearlet transform to classify colon lesion. The MFF-CNN based on shearlet transform can realizes automatic feature extraction, feature fusion and classification for colon cancer, which have an accuracy of 96%. The superior performance of the proposed method lays a foundation for the application of deep learning in medical image identification, which is opening a new perspectives for real-time, objective and accurate diagnosis of cancer

METHODOLOGY

SHEARLET TRANSFORM

Database The database used in this paper is Hematoxylin and

Feature extraction

CONCLUSION